Creating Your First Agent
Name the business, teach the agent, and try a private question before customers can hear it
Creating Your First Agent#
An agent answers from your business, not from a generic script. Setup is private until you confirm a live channel.
| Step | Where | Ready when |
|---|---|---|
| Create the agent | Project dashboard → New AI agent | You are in the intended project and agent |
| Name the business | The setup flow after the first agent | The business name and what you offer are saved |
| Teach the agent | Agent Studio → Data sources | Current facts are added, and training has finished where it is required |
| Try a private question | Playground and Agent Studio → Test | A real customer question gets an answer you trust |
| Launch | Deploy | A channel shows a confirmed live activation, and a controlled customer message gets the intended reply |
Finishing setup, connecting an account, or approving a private reply does not turn the agent on for customers.
Create the agent#
- Open the project dashboard.
- Choose New AI agent.
- Enter an Agent Name and choose Create Agent.
The first agent continues into setup and asks for the business or organization name. Later agents open the agent workspace. You can rename the agent in Settings → General. The website's first message is separate: set Initial message under Deploy → Chat widget or Chat bar → Content.
If creation sends you to Settings → Plans, the project has reached its agent limit or access is locked. Compare the live plan there instead of using an old price list.
Teach it#
Open Agent Studio → Data sources and choose Add source. Start with hours, services, prices, policies, and common questions.
Adding or uploading a source does not by itself make it answerable. Files, text, Q&A, websites, and Google Drive need the Train agent action when the page offers it. The button says Retrain agent after the first successful training. Structured data, time-sensitive content, and social posts are prepared in their own source workspace.
Review processing errors before you rely on the content. See Data sources and Training your agent.
Try it privately#
Ask a question a customer would actually send.
- Playground is a private chat. Use the preview channel to check Web Widget or Messenger. Customers cannot see this thread.
- Agent Studio → Test saves answer checks and simulations you can run again after you change sources or guidance.
Do not treat a good private answer as launch.
Turn it on#
Open Deploy, finish the channel you intend to use, and confirm live replies. On website chat, that switch is AI Auto-Reply. Turning the widget or bar on, or saving its look, does not answer customers by itself. Then send one controlled customer message and read it in Activity → Inbox.
Tone, model, and creativity live in Agent Studio → Guidance. The model list shows the current credit cost. Tools, Procedures, and Workflows are for actions such as booking or payments. A policy document does not connect those services.
If answers are weak#
| What you see | What to check |
|---|---|
| Generic or "I don't know" | The fact is missing, still processing, or not trained |
| Old information | Edit the source that should win, then retrain if the page asks for it |
| Right words, wrong action | Test the tool result, not only the sentence |
| Customers heard a draft | Live replies were enabled before the private check |